Hybrid Retrieval-Augmented Generation for Scientific Knowledge Discovery: Toward Grounded, Cross-Disciplinary Synthesis in the Era of Large Language Models

Abstract

The proliferation of large language models has intensified interest in systems capable of grounding generative output in verifiable, up-to-date scientific evidence. This paper examines hybrid retrieval-augmented generation, an architectural paradigm that fuses sparse lexical retrieval with dense semantic retrieval, as a mechanism for supporting knowledge discovery across disciplinary boundaries. The study is motivated by the persistent problem of hallucination in generative systems and by the observation that neither purely lexical nor purely neural retrieval adequately serves the terminological precision and conceptual abstraction that characterize scientific literature. Employing a theoretical and comparative conceptual methodology, the paper synthesizes findings from information retrieval, natural language processing, and science-of-science scholarship to construct an integrative framework in which hybrid retrieval, knowledge-graph contextualization, and citation-grounded generation operate in concert. The analysis draws on a structured review of foundational and contemporary peer-reviewed sources to evaluate how hybrid architectures address recall gaps associated with rare technical vocabulary while preserving the semantic sensitivity required for cross-domain conceptual matching. Findings indicate that hybrid configurations consistently outperform single-mode retrieval baselines on recall and downstream factual accuracy, particularly within interdisciplinary and terminology-dense corpora, though at increased computational and architectural cost. The paper concludes that hybrid retrieval-augmented generation constitutes a meaningful, though not yet fully mature, contribution to trustworthy scientific automation, and it proposes directions for evaluation standardization and governance that would strengthen its adoption within research infrastructures.